A hybrid approach for real-time respiratory motion prediction for radiotherapy applications

Predicting tumor motion in motion adaptive radiotherapy is challenging due to the irregular and non-stationary nature of respiratory motion. Existing methods often suffer from large prediction errors caused by time-varying irregularities and intra-trace variabilities of respiratory motion, and high computational time requirements. To overcome these issues, we propose hybrid real-time framework called BMFLC-EMD-RVFL, which integrates bandlimited multiple Fourier linear combiner with Kalman filter (BMFLC-KF), empirical mode decomposition (EMD), and random vector functional link (RVFL) with incremental learning. The BMFLC-KF algorithm extracts the respiratory motion weights, which are decomposed into intrinsic mode functions (IMFs) and residues using EMD. RVFL predictors are trained for these IMFs and residues, and their aggregated prediction results formulate the BMFLC predicted weights. These weights are then multiplied by the known reference vector containing sine and cosine components of predefined input frequencies to formulate predicted respiratory motion signal. We evaluated our method on 304 respiratory motion traces from 31 patients, covering various prediction lengths. The results demonstrate that the BMFLC-EMD-RVFL framework delivers superior prediction performance and reduced computational time compared to existing methods.